{
 "schema": "cain.e127.cv_selection.v1",
 "at": "2026-11-21T04:42:24Z",
 "entity": "FIRME AI CORPORATION",
 "experiment": "E127-A cross-validation hypothesis selection",
 "preregistration": {
  "schema": "cain.e127.preregistration.v1",
  "frozen_at": "2026-11-21T04:41:18Z",
  "title": "E127-A \u2014 cross-validation hypothesis selection to fix the E125-A replication failure",
  "diagnosis": "E125-A showed the E124 full arm (85.5%) losing to a compute-matched random search (92.5%) on fresh variants. The failure mode is selection: several depth-2 hypotheses fit the 12 train examples exactly, and fixed-order search keeps a spurious one.",
  "intervention": "Among hypotheses from the SAME enumeration space, fit on the first 8 train rows, select by (fits all 8, accuracy on the held-back 4 train rows, shorter spec). No new hypothesis class, no sealed data used for selection.",
  "splits": {
   "development_variants": [
    7,
    8,
    9
   ],
   "sealed_test_variants": [
    15,
    16,
    17,
    18,
    19
   ],
   "excluded": "0-6 used by E124; 10-14 used by E125-A and therefore not reused as a test set"
  },
  "arms": [
   "baseline_current",
   "compute_matched",
   "cv_filter"
  ],
  "primary_endpoint": "sealed accuracy on variants 15-19 (200 instances) for cv_filter vs baseline_current",
  "success_threshold": "cv_filter beats BOTH baseline_current and compute_matched by >= 5.0 percentage points, exact McNemar p < 0.05 vs baseline_current",
  "failure_rule": "if the gain is below +5.0 pts or p >= 0.05, report NEGATIVE and do not promote",
  "budget": "local CPU only, $0",
  "limits": [
   "constructed generator",
   "same eight paradigms",
   "no external benchmark",
   "no human review"
  ]
 },
 "development": {
  "baseline_current": 0.8083333333333333,
  "compute_matched": 0.8416666666666667,
  "cv_filter_400": 0.825,
  "cv_filter_full": 0.9833333333333333,
  "cv_vs_baseline_mcnemar": {
   "cv_only": 22,
   "baseline_only": 1,
   "p": 5.7220458984375e-06
  },
  "n_instances": 120
 },
 "sealed_test": {
  "baseline_current": 0.805,
  "compute_matched": 0.86,
  "cv_filter_400": 0.82,
  "cv_filter_full": 0.995,
  "cv_vs_baseline_mcnemar": {
   "cv_only": 38,
   "baseline_only": 0,
   "p": 7.275957614183426e-12
  },
  "n_instances": 200
 },
 "delta_vs_baseline": 0.18999999999999995,
 "delta_vs_compute_matched": 0.135,
 "verdict": "POSITIVE_PROMOTABLE",
 "threshold": ">= +5.0 pts vs BOTH and McNemar p<0.05 vs baseline",
 "limits": [
  "constructed generator",
  "same eight paradigms",
  "no external benchmark",
  "no human review",
  "cv_filter uses more hypothesis evaluations than the 400-step controls"
 ],
 "agi_status": "NOT ACHIEVED"
}
